Re: [RFC PATCH net-next v1 4/4] net: page_pool: use netmem_t instead of struct page in API

[Mina Almasry]

On Thu, Jan 4, 2024 at 12:48 AM Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
>
> On 2024/1/4 2:38, Mina Almasry wrote:
> > On Wed, Jan 3, 2024 at 1:47 AM Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
> >>
> >> On 2024/1/3 0:14, Mina Almasry wrote:
> >>>
> >>> The idea being that skb_frag_page() can return NULL if the frag is not
> >>> paged, and the relevant callers are modified to handle that.
> >>
> >> There are many existing drivers which are not expecting NULL returning for
> >> skb_frag_page() as those drivers are not supporting devmem, adding additionl
> >> checking overhead in skb_frag_page() for those drivers does not make much
> >> sense, IMHO, it may make more sense to introduce a new helper for the driver
> >> supporting devmem or networking core that needing dealing with both normal
> >> page and devmem.
> >>
> >> And we are also able to keep the old non-NULL returning semantic for
> >> skb_frag_page().
> >
> > I think I'm seeing agreement that the direction we're heading into
> > here is that most net stack & drivers should use the abstract netmem
>
> As far as I see, at least for the drivers, I don't think we have a clear
> agreement if we should have a unified driver facing struct or API for both
> normal page and devmem yet.
>

To be honest I definitely read that we have agreement that we should
have a unified driver facing struct from the responses in this thread
like this one:

https://lore.kernel.org/netdev/20231215190126.1040fa12@xxxxxxxxxx/

But I'll let folks correct me if I'm wrong.

> > type, and only specific code that needs a page or devmem (like
> > tcp_receive_zerocopy or tcp_recvmsg_dmabuf) will be the ones that
> > unpack the netmem and get the underlying page or devmem, using
> > skb_frag_page() or something like skb_frag_dmabuf(), etc.
> >
> > As Jason says repeatedly, I'm not allowed to blindly cast a netmem to
> > a page and assume netmem==page. Netmem can only be cast to a page
> > after checking the low bits and verifying the netmem is actually a
>
> I thought it would be best to avoid casting a netmem or devmem to a
> page in the driver, I think the main argument is that it is hard
> to audit very single driver doing a checking before doing the casting
> in the future? and we can do better auditting if the casting is limited
> to a few core functions in the networking core.
>

Correct, the drivers should never cast directly, but helpers like
skb_frag_page() must check that the netmem is a page before doing a
cast.

> > page. I think any suggestions that blindly cast a netmem to page
> > without the checks will get nacked by Jason & Christian, so the
> > checking in the specific cases where the code needs to know the
> > underlying memory type seems necessary.
> >
> > IMO I'm not sure the checking is expensive. With likely/unlikely &
> > static branches the checks should be very minimal or a straight no-op.
> > For example in RFC v2 where we were doing a lot of checks for devmem
> > (we don't do that anymore for RFCv5), I had run the page_pool perf
> > tests and proved there is little to no perf regression:
>
> For MAX_SKB_FRAGS being 17, it means we may have 17 additional checking
> overhead for the drivers not supporting devmem, not to mention we may
> have bigger value for MAX_SKB_FRAGS if BIG TCP is enable.
>

With static branch the checks should be complete no-ops unless the
user's set up enabled devmem.

> Even there is no notiable performance degradation for a specific case,
> we should avoid the overhead as much as possible for the existing use
> case when supporting a new use case.
>
> >
> > https://lore.kernel.org/netdev/CAHS8izM4w2UETAwfnV7w+ZzTMxLkz+FKO+xTgRdtYKzV8RzqXw@xxxxxxxxxxxxxx/
>
> The above test case does not even seems to be testing a code path calling
> skb_frag_page() as my understanding.
>
> >



-- 
Thanks,
Mina